Point Shaving in the NBA: An Economic Analysis of the National Basketball Association s Point Spread Betting Market

Transcription

1 Point Shaving in the NBA: An Economic Analysis of the National Basketball Association s Point Spread Betting Market May 11 th 2007 Jonathan Gibbs jmgibbs@stanford.edu Department of Economics Stanford University, Stanford, CA Advisor: Roger Noll Abstract In this paper, I examine the point spread betting market of the National Basketball Association for indications of cheating through point shaving. Previous studies have illustrated the contradictory phenomenon of simple, profitable betting rules persisting in a market of informed traders. While normally dismissed due to the salaries earned by professional players, I propose point shaving as an explanation for this persistent market inefficiency. Through analysis of the distribution and symmetry of the forecast errors, I arrive at a rough approximation of point shaving s prevalence. In studying in game scores, I find conclusive evidence that the point spread influences game outcomes and has causative effects on the probability of the favored team covering the point spread. The key economic concept these results support is that with the non linear payout structure of the point spread betting market creating a discontinuity over the value of final game margins, individuals are altering their behavior in response to these distorted incentives. Keywords: sports betting, gambling market, NBA, sports book, efficient markets, wager I am indebted to Roger Noll and Nick Bloom for their insights and patience throughout this process. I would like to thank my draw mates for helping to shape and listening to my ideas, and my parents, grandparents, sister and girlfriend for their help, support, love, and encouragement. All errors are my own.

2 Jonathan Gibbs 2 May I. Introduction Sports gambling markets have long been used by economists to study and test hypotheses about the operation of financial markets. Among the useful attributes of betting markets is that they have a definitive endpoint at each game s conclusion, yielding an objective value of the traded asset. Point spread betting markets also provide predictions, in which prices represent forecasts of the final point differential of a game. Data on betting lines and game outcomes have been used in recent empirical financial research to study questions such as how markets incorporate information and the factors affecting price variations. A literature review indicates that the overall performance of the National Basketball Association (NBA) betting market is efficient, with public and private information incorporated into prices. This paper focuses on a segment of the NBA betting market that may be susceptible to the equivalent of insider trading. For basketball, this phenomenon manifests itself as point shaving. In return for a bet s monetary payoff, a point shaving player chooses to affect his play negatively in order for his team to still win, but by a smaller margin than the predicted point spread. When discussing the possibility of point shaving in the NBA, the overwhelming majority of basketball fans immediately point to the large contracts players are paid as evidence against this form of cheating. In ESPN.com writer Chris Sheridan s blog post of February 14, 2007, he recalls asking NBA commissioner David Stern, if the potential for corruption -- players fixing games or shaving points -- was a significant concern of his. [Stern s] answer was no. In their analysis of the NBA betting market, Dobra, Cargill, and Meyer (1990) succinctly stated that although [they] recognize that fixes are not impossible to arrange,... they are very unlikely, and therefore, the [betting] lines... are probably not substantially affected by this kind of activity (p.247). It is true that the NBA has become a billion dollar global business and that for

3 Jonathan Gibbs 3 May 11, 2007 the season, players averaged an annual salary of $5.215 million, a 1,475% 1 increase since the season (Ford 2006). However, simply because an individual already earns a high wage, does not imply that the potential to earn additional money will not be attractive. There is a disincentive associated with the potential for lost wages if a player is caught shaving points, however two other factors influencing player salaries are not often mentioned. First, the demand for NBA contracts far exceeds their supply. The total number of NBA roster spots is fixed at a maximum of only 450, (thirty teams each with fifteen players). The combination of a large player pool and few available contracts engenders substantial market competition, with past NBA experience not close to a guarantee of future earnings. Second, even the most competitive and talented players are not assured of long term NBA success 2. Basketball is a physically and mentally demanding sport, with injuries almost inevitable. Even small medical issues can bring a player one step closer to retirement, implying these players world class skill only has earnings potential for a limited duration. Aside from the risk of injuries, this future earning potential is further negatively impacted by expected declines to a player s skills resulting from aging. The sharp decrease in expected wages accompanying players retirements emphasizes the limited window of opportunity. The uncertainty created by these two factors works to somewhat counteract the decreasing marginal utility of money argument described above. There is also anecdotal evidence suggesting players already influence their play for financial gain. NBA players constantly face accusations of playing to accumulate individual statistics instead of playing to elevate the team s chances of winning. This me first attitude designed to capitalize on the dollars available on the free agent market further separates the player s goals from that of the team. The contract year phenomenon provides similar

4 Jonathan Gibbs 4 May 11, 2007 anecdotal evidence suggesting that players in the final year of their contract outperform their past statistical averages in order to earn a larger contract at year s end. Once signed to a new contract, players then revert to their pre contract year level of production. This boost in performance and subsequent regression demonstrates the me first attitude, as teams should expect players to always perform at their peak, and demonstrates players altering their level of play in response to monetary incentives. Similarly responding to large monetary incentives, in 2007, the NBA held its annual all star game in Las Vegas for the first time. While one major stipulation for having the game in Las Vegas was that all sports books could not offer gambling action on the game, NBA players and gambling have gone hand in hand for a long time 3. ESPN.com columnist Bill Simmons in his June 19, 2006 article for ESPN The Magazine goes so far as to claim that a predisposition to gambling is a positive trait within a basketball player. The intense desire to compete and more importantly the overwhelming need to win characterize the best basketball players, separating them from the rest of the league. He posits that the fortitude to put everything on the line and take risks in an effort to win is as much a necessity in basketball as it is in high stakes gambling. Michael Jordan, arguably the best basketball player ever, was a notorious gambler during his career. There are a multitude of stories with regards to Jordan and his gambling, most often concluding, similarly to Simmons, that his gambling demonstrates his legendary drive to compete and all consuming need to win. The similarity in requisite instincts and skills between the two activities makes the transition from basketball to gambling seamless. The question thus becomes, have players crossed the line, and begun betting on NBA games? Sports gambling has enjoyed a renaissance in the past fifteen years with the internet creating an explosion of betting sites offering daily NBA betting lines, or point spreads. No

5 Jonathan Gibbs 5 May 11, 2007 longer restricted to the sports books of Las Vegas, annual betting volumes in the millions are placed online and around the world. Bookies attempting to pay off players to fix games are obsolete as players can with far more ease simply place a bet themselves. This creates a dangerous intersection as the game making these players rich is also a bettor s haven. On March 20, 2007, in a road loss against the Portland Trailblazers, all star guard Gilbert Arenas of the Washington Wizards was involved in a betting incident. He made separate bets with two different fans during the game for $10, in which if he had the opportunity to take a game winning shot, he would win the bet if he made that final shot; Arenas got his chance, missed, and paid up (Associated Press 2007). He was quickly admonished by the NBA when the bets were made public after the game, but it furthers speaks to the connection between basketball players and gambling. These players are young, hyper competitive, rich, attracted to gambling and armed with better information over NBA game outcomes than the odds makers themselves. As these players have superior knowledge of the sport and more importantly, a sphere of influence over game outcomes, simply setting a betting line provides NBA players with incentives to cheat. II. The NBA Betting Market The NBA betting market needs to be addressed before asking the question of how to detect whether NBA players are cheating for monetary gain. For any NBA game a variety of factors affect each team s chances of winning, including the teams relative strengths, which team is at home, the results of each team's previous games, luck, as well as a host of other elements. Based on an analysis of these factors, one team usually has a greater chance of winning. Thus, in order for individuals to bet on one team or the other, odds makers have two options for creating a betting market. They can either decrease the payout for those betting on the team more likely to win, or they can handicap each team s chances of winning by

6 Jonathan Gibbs 6 May 11, 2007 announcing a betting line. Both options exist within the betting market with the former being less relevant as the latter is much more prevalent. This analysis limits its focus to solely the betting line market as the discontinuity in its payout structure discussed in the next section provides the incentive for players to cheat. As an example, for the NBA game on March 18, 2007 between the Detroit Pistons and the Dallas Mavericks, the betting line was set at the Mavericks being favored by 5.5 points. In this instance, a bet on the favorite, the Mavericks, wins if the Mavericks win by more than 5.5 points; conversely, a bet on the underdog, the Pistons, wins if the Pistons win OR if the Mavericks win by less than 5.5 points. Since NBA scores increase by integers, while betting lines are set in 0.5 point increments, this betting line was guaranteed to produce winning and losing bets. If instead the Mavericks had been favored by 6 points and the outcome of the game was such that the Mavericks won by exactly 6 points, the bet would have pushed, and bets for both sides would have been returned in full. The question of how the point spread is set is interesting. The Las Vegas Sports Consultants (LVSC), a company owned by Michael Roxy Roxborough establishes the odds for approximately 75% of the licensed sports books in Nevada, with most illegal and online books in and out of Nevada drawing their odds from what the LVSC develops (Stern 1999). These betting lines are posted at 8:00 AM Pacific Standard Time (PST) on game day and fluctuate from market trading until each game s tip off, at most eleven hours later. The company s stated goal and the generally accepted assumption is that these opening betting lines are set to equalize the amount of money bet on each side of the betting line, minimizing the bookmakers financial risk and guaranteeing their revenue. (Whether this assumption holds is discussed in Section IV, the literature review.) This riskless profit is assured by the eleven for ten trading rule where a

7 Jonathan Gibbs 7 May 11, 2007 standard fee (the vigorish) is taken from all winnings; the price the bookmakers charge for creating the betting market. By taking in 100% of the money from losing bets while only paying out 90.91% of the money from winning bets, as long as there is an equal amount of money bet on either side of the betting line, the bookmaker profits regardless of the game s outcome. By attempting to balance the dollar amount bet on each team, if the NBA betting market is efficient, the point spread corresponds to the number of points by which the favored team is expected to win. If the opening betting line does not attract equal monies bet on either side of the betting line, the sports book often moves the point spread during the eleven hour period between opening and closing to avoid risk. Unlike in pari-mutuel betting where odds are recalculated as bets are placed, point spread bets lock in the betting line at the time of the bet s placement. Thus if the Mavericks were underpriced at their opening betting line (only favored by 3 points, when the optimal point spread is 5.5 points), there will be excess demand for the Mavericks and insufficient demand for the Pistons. This deviation from the optimal betting line signals a profit opportunity to the betting public at the expense of the bookmaker, and would induce betting line changes in response to market activity. The closing betting lines therefore reflect the dollar weighted average of the public s opinion on the outcome of the game, not the bookmakers. The point spread can thus be viewed as a market clearing price and a forecast of a game s final score, with the difference between the actual outcome and the betting line s prediction interpreted as a forecast error. If NBA players are cheating, negatively affecting their play to win bets, these forecast errors could demonstrate an observable bias relative to the point spread.

8 Jonathan Gibbs 8 May 11, 2007 III. Point Shaving Conceptually There are three methods through which players could influence their play in order win outside wagers. The first and most prominent method is by throwing games. In baseball, eight players on the Chicago White Sox of 1919 took bribes during the World Series in return for purposefully losing games. Similarly, NBA players could negatively alter their play to this level inducing their teams to lose strategic games on which they have placed bets. However, the poor level of play necessary to pull off this feat could have problematic ramifications for the player. First, the player could be benched by the coach due to his poorer performance, preventing his influence on the game s outcome. Second, his poor play could impact future contracts and thus future wages. Third, part of the reason players are attracted to gambling and placing bets is their hyper competitive nature, which would go unfulfilled resulting from their team s loss. These three aspects make the possibility of throwing games unappealing. The second method by which players could alter their level of play for monetary gain is by exceeding the player s expected effort level. If the betting public expects a player to play at a certain level of intensity and sets the betting line accordingly, the player could choose to exceed that effort level, having placed a bet on his own team to win and cover the point spread. The extra unexpected effort the player exerts would be unaccounted for in the betting market and would make his bet s outcome more likely. This theory is possible and could be demonstrated in the data; however it does not seem probable. First, players already have strong monetary incentives to exert as much effort as possible in order to earn as large a contract as possible. Second, after increasing one s effort for a specific game, there may be negative effects from the coach or fans when in subsequent games the player s effort returns to previous levels. Third, it is

9 Jonathan Gibbs 9 May 11, 2007 much less appealing to increase one s effort, when a slight decrease in effort to instead shave points would have the same monetary result. The third method, the practice of point shaving, exploits the discontinuity in the betting line s payout structure. While the betting line is designed to balance the betting market, it does not correspond to the incentives of the teams: the teams only care about wins and losses. This asymmetry creates a window of space in which the favored team can still win the game while a bet on the favored team loses. This window is an opportunity for players to shave points without adversely affecting their teams wins and losses. If done carefully, a player on a favored team could slightly reduce his effort to change the expected final margin without changing his team s likelihood of winning. This decrease in effort, done in conjunction with the placement of a bet on the team s opponent to cover the point spread, is point shaving. In this scenario, the player gets the best of both worlds; his team wins the game and he receives a financial payoff as a result of purposely slightly underperforming. As shown in Figure 1, if the player is exerting expected maximum effort, the probability of just failing to cover the point spread (B 1 ) equals the probability of just covering the point spread (C 1 ). However, by slightly reducing his effort he can induce the game to end up in (B 1 ) instead of (C 1 ) without changing the outcome of the game, solely shaving points off the final margin of victory. In essence, the betting market assigns values to final margins with a discontinuous jump in value occurring at the point spread. As the final margin has little if not no intrinsic value, (other than determining whether the team wins or loses), the betting market s discontinuous valuation of final game outcomes provides incentives to arrive on one side of the betting line relative to the other. Thus, as one s effort is more easily decreased than increased,

10 Jonathan Gibbs 10 May 11, 2007 point shaving is the response to these monetary incentives fundamentally created from the setting of the betting line. Once again, the typical response to the possibility of point shaving is that the large salaries NBA players earn should counteract the incentive to cheat for monetary gain. However, as was discussed above and through further breaking down the decision making process in equations (1) and (2), this argument does not seem satisfactory. Benefits: (Point Shaving Success Rate)*(Betting Payoff) (1) Costs: (Lost Future Income + Lost Reputation)*(Probability of Detection) (2) In comparing the two equations, since the (Probability of Detection) is near zero, the expected costs of point shaving are tiny relative to the expected benefits: (1) > (2). Additionally, by selecting only certain games in which to engage in point shaving the player is able to maximize the (Point Shaving Success Rate), while further minimizing the risk of detection. Games in which the player s team is heavily favored maximize the window in which the team can win, while a bet on the underdog to cover pays off, or in returning to Figure 1, maximizes the area in region B. Thus, while each team s overriding goal is to win each game, the incentive to shave points is greatest for games in which one team is heavily favored. If point shaving is occurring, this theory suggests that it is in these games the practice would most likely be observed. At the college basketball level, where players are not paid, there have been three major point shaving, cheating, scandals in the past twenty five years, two occurring in the past ten years. Recent statistical analyses of past college game results indicate that this point shaving problem may be more serious and widespread than is generally believed (Wolfers 2006). Yet, despite college players making the annual transition to the NBA, there has never been a point shaving scandal in the history of the NBA. Conventional wisdom ascribes the hefty salaries

11 Jonathan Gibbs 11 May 11, 2007 NBA players are paid as sufficient to eliminate any temptation to shave points for financial gain. However, the NBA s expensive culture, the players limited and uncertain earning potential, and the players ultra competitive nature, combined with the miniscule likelihood of being caught may counterbalance this mainstream thinking. Furthermore, even if personal (or friend/familial) financial motivations are absent, the overpowering need to win or need to satisfy one s ego, may still make betting on NBA games attractive. The focus of this analysis is to detect whether NBA players have in fact crossed the line and begun betting on their own games. Utilizing data over betting lines, in game scores, and game outcomes, I propose to investigate the suspicion that gambling has caused NBA players to affect their play to influence the outcome of NBA games in order to win bets. If this phenomenon is occurring, it should leave observable anomalistic footprints in the data. Thus, the question asked by this empirical study is whether there are statistical indications of past point shaving in the NBA. This paper proceeds with a discussion of the relevant literature and their examinations of the NBA betting market including a discussion of how this analysis shall enhance the existing literature and attempt to answer the described question above. The next section details the econometric and statistical methods used as well as the data. An in depth discussion of the results follows before concluding remarks. IV. Literature Review The majority of the scholarly literature discussing this subject focuses on (1) revealing the subtle inefficiencies in a mostly efficient NBA point spread betting market and then (2) attempting to provide an explanation as to why these inefficiencies persist. It has become quite well established that this market generally does an excellent job at inducing betting lines to accurately predict game outcomes. The debate thus centers on the accompanying notion that

12 Jonathan Gibbs 12 May 11, 2007 simple profitable betting strategies seem to endure. Overall market efficiency predicts the demise of these strategies over time as experienced bettors exploit the profit making opportunities. It is the conjecture of this analysis that these unexplained deviations from efficiency would be consistent with the result of point shaving in the NBA. The first research into the efficiency of the basketball betting market performed by Camerer (1989) built upon the psychological misperception of the hot hand, and how individuals incorrectly judge the randomness of sequences. The cognitive psychology paper by Gilovich et al. (1985) experimentally demonstrated that while subsequent shots by college basketball players are uncorrelated, bettors tend to believe the notion that a player is more likely to make (miss) his next shot given that he has made (missed) his previous. This result was attributed to one s tendency to reject a lengthy sequence of makes or misses as part of a random distribution, albeit that the probability of encountering such a sequence is quite high. Camerer extended this hot hand fallacy to see if the basketball betting market could accurately adjust betting lines for teams on winning or losing streaks. He aggregated game result and betting line data across three seasons for teams on winning and losing streaks. He then computed the streaking team s probability of beating the point spread in their next game given their current streak length. This data showed that the betting market is able to accurately account for games played by teams on winning streaks, while for teams on losing streaks the probabilities of beating the point spread significantly differ from (0.5). He noted that for teams on losing streaks, these probabilities fall into the profitability range for bettors, although just barely. He concluded that for basketball betting, the hot hand fallacy exists but it is slight, (p.1260) as traders can study past data and exploit such errors easily (p.1260).

13 Jonathan Gibbs 13 May 11, 2007 One large problem with Camerer s piece is that it started with the assumption that the hot hand is a myth and simply worked to show that the betting market does not adequately adjust for streaking teams. By only measuring the winning frequency of bets against the point spread, Camerer limited himself by his design. Brown and Sauer (1993a) addressed this issue specifically, instead offering the hypothesis untested by Camerer that simultaneous adjustments occur in a team s performance level and in market point spreads during streaks. They created a point spread pricing model to estimate the baseline expected amount by which one team would be favored over another and then tested whether this value changes based on the relative streaks of the two teams. Utilizing data over six seasons, twice that of Camerer, Brown and Sauer tested (H1) whether the hot hand is irrelevant to both the betting market and game outcomes, (H2) whether people believe in a non existent hot hand effect as posited by Camerer, and (H3) whether point spreads take into account actual hot hand effects. Their results definitively rejected H1 that streaks are irrelevant to the betting market and to team performance. However they could not reject either H2 or H3 that either the hot hand is a myth or that recent previous performance affects immediate outcomes. Brown and Sauer convincingly demonstrated that point spreads react to streaks, increasing point spreads by 0.25 points for teams on two game winning streaks and by 0.67 points for teams on four game winning streaks. However, the reason why was inconclusive as they were unable to tease out whether the betting market was reacting to game fundamentals or to noise, which in turn was providing for the profitable strategies. This question of fundamentals or noise was the motivation in the next paper by Brown and Sauer (1993b), retesting their three hypotheses in a more generalized form. Utilizing their point spread pricing model, they demonstrated that the explanatory power of the point spread is

14 Jonathan Gibbs 14 May 11, 2007 quite high (R 2 = 0.889), however there still remains a nontrivial unpredictable component. Brown and Sauer applied this model to predict games out of sample by using data over the two teams previous twenty games to predict its next five. Even out of sample, this model explained more than 85% of game to game variation in point spreads. Additionally, less than a quarter of the out of sample forecast errors were more than two points away from zero implying the distribution of the forecast errors were tightly concentrated about zero. From testing actual score differentials, actual point spreads, and predicted point spreads they were able to once again conclusively reject H1, that the unpredictable component is pure noise. However, their tests could not distinguish between simple and strong fundamentals comprising this unknown component. Simply put, relevant information is embedded in this unpredictable element; however, it is not strong enough to be able to be used to distinguish between close point spreads, as in 1.5 versus 2. Furthermore, while this component is small relative to its more rigorous model counterpart, Brown and Sauer found that without it, point spreads would be significantly biased, the noise component being essential to predicting the outcome of the game itself (p.1208). Fortunately, one major piece missing from the above studies, the possibility of serial correlation in game to game point spreads, was tested for by Oorlog (1995) and found to be irrelevant. While the above papers tested whether by the end of the season the point spread was an accurate pricing mechanism, they did not test to see if it was an accurate pricing mechanism within the season. Brown and Sauer s positive out of sample results provide the intuition for why this eventuality was not significant, however without explicitly testing for serial correlation in the point spread between games, the analysis was not complete. By testing the regression residuals, Oorlog was able to reject autocorrelation at the 5% significance level for 51 of the 54

15 Jonathan Gibbs 15 May 11, 2007 team seasons tested. His results express that there is no such idea as momentum for or against the point spread as the wagering public accurately takes in new information, utilizing it to set future point spreads. These results support the hypothesis that the betting market uses past information to improve future predictions. Together, these four papers arrive at similar conclusions. They demonstrate that (1) the point spread is an accurate predictor of game outcomes. Then, they show that (2) the point spread reacts to relevant information to improve its game to game forecast. Next, they reveal that (3) a mix of fundamentals and noise make up the forecast errors. Finally, they conclude that (4) the efficient market should eliminate simple profitable betting strategies as the fundamentals comprising the unpredictable component become exploited and known. At that point the pivotal piece of literature explicitly and rigorously studying the efficiency of betting lines was authored by Gandar et al. (1998). This paper demonstrated that the betting public fundamentally consists of informed traders creating the influential market forces necessary to remove biases/inefficiencies from betting lines. Instead of taking the betting line as a singular value for a specific game as was done in the previous literature, Gandar et al. gathered data over each game s opening and closing betting lines to test the market s ability to trade to a more efficient point spread. They first noted that during the eleven hour market period between opening and closing betting lines, approximately 80% of NBA games undergo betting line changes with nearly 0.33 being greater than a point in magnitude. Thus bookmakers often respond to betting imbalances by moving the point spread, in an attempt to attract more betting on the less popular position, in order to balance the dollar amount bet on either side. The authors provided four possible reasons for betting imbalances in the market: (1) release of new

16 Jonathan Gibbs 16 May 11, 2007 information after betting lines are announced, (2) randomness of bettor order, (3) prediction errors on the part of the betting public, and (4) prediction errors on the part of the bookmaker. The first two possibilities were dismissed as unable to explain all of the variation between opening and closing betting lines. Information release (1) did not seem to fit as the period of market activity was too short to adequately explain all market betting line changes. Order flow randomness (2) seemed implausible as large betting line changes require continued betting on one side even as that side becomes less and less favored. The remaining two possibilities provided readily testable theories. Gandar et al. first tested the relative forecast accuracy of opening and closing betting lines and found closing betting lines to be statistically superior predictors of game outcomes. This first result supports the notion that market activity improves the ability of the betting line to project final scores, exploiting biases in bookmaker s initial betting lines. However, the actual differences between the two point spread sets were quite small: not large enough to conclude that informed bettors dominate the market. More importantly, what matters to bettors is not by how much a team beats the point spread (the relative accuracy of the betting line), but simply whether their team is able to exceed it. This led Gandar et al. into the bulk of their testing, examining the magnitude and direction of betting line changes to discover if the market driven point spread movements contain information relevant to game outcomes or if they are instead noise. The results they found are conspicuous. For every individual positive betting line change the actual team winning proportion of bets at opening betting lines is always greater than (0.5). Similarly, the converse is true for negative betting line changes. Additionally, for the 2,036 games in which the betting line did not change, the probability the favored team went on to beat the point spread did not significantly differ from (0.5). Simply put, betting lines move towards

17 Jonathan Gibbs 17 May 11, 2007 the point where each team s likelihood of covering the point spread is (0.5), effectively removing biases from opening betting lines. Furthermore, when betting lines do not move, opening betting lines are already at this point of equality. As for the magnitude of the change in the point spread, Gandar et al. discovered that the larger the betting line change the larger the distortion from (0.5) in the initial point spread. These results led to the conclusion that in almost all instances bettors are able to move lines by a magnitude sufficient to remove opening line biases by the close of betting, (p ) and that informed traders are both present and influential in this market (p.398). The data decisively demonstrated that bookmaker s unbiased opening betting lines are not moved through bettor activity, while bookmaker s biased opening betting lines are exploited by the informed betting public 4. Thus, this paper concurs with Brown and Sauer s (1993b) result that bookmakers incorporate most if not all of the fundamental information into opening betting lines, however this study pushed further, conclusively showing that the betting public processes privately held fundamental information to yield improvements in prices. In conjunction with the findings of Gandar et al., Paul and Weinbach s (2005) analysis of the NBA betting market brings this literature to a crossroads. We know that the bookmakers have strong financial incentives to set unbiased betting lines and that the betting public induces betting line adjustments by exploiting biased point spreads. Furthermore, we know that closing betting lines reflect both public and private information yielding an efficient betting market. Yet, Paul and Weinbach demonstrate using closing betting lines that while the market as a whole is efficient, simple profitable betting strategies still exist. They found that the strategy of betting double digit underdogs rejects the null hypothesis of a fair bet, and the strategy of betting double digit home underdogs even rejects the null hypothesis of no profitability. Given both the readily available volumes of past game data and the informed bettors present in the market

18 Jonathan Gibbs 18 May 11, 2007 consistently trading to efficient prices, why does this market fail to exploit and eliminate this profitable betting strategy across seven seasons of data? Paul and Weinbach suggest four different possibilities for the market s failure to accurately price games with large underdogs. The first two suggestions that (1) bettors may overestimate the relative strengths of the best teams compared with the worst teams or that (2) players or coaches may shirk in lopsided contests, not putting forth maximum effort, do not seem credible. It has already been shown that the betting public learns from past game results incorporating information into future prices. The simple explanation of overestimation of relative team abilities seems contrived and unlikely. As for players shirking, it may very well be true that in lopsided contests the final margin of victory decreases in the final minutes as a result of decreased effort by the winning team. However, the decrease in effort by teams winning by large margins is something for which the market should be able to account. There would have to be a systematic bias of underestimating the decrease in effort by the leading team for this possibility to explain why large underdogs tend to cover too often. Otherwise, the expected decrease in effort would yield an expected point value by which the final margin decreases, which would then be integrated into the betting lines. I plan on further examining this possibility as while the narrowing of the final score should be independent from the initial point spread solely affected by the margin as the game winds down, if players are shaving points this would not be the case. The remaining two possibilities Paul and Weinbach present, that (3) gamblers may receive a disutility from betting on large underdogs or that (4) betting limits may restrict the ability of informed traders to discipline the market are more intriguing. It is plausible that bettors may find the idea of rooting for a team to lose by less than 18 unsavory. Note however

19 Jonathan Gibbs 19 May 11, 2007 that this is not the observed bias towards longshots found in pari-mutuel horse race betting as the betting line is set to make the chance of each team covering the point spread equal. Thus, this suggestion purports that these informed bettors are leaving profits unclaimed, making this conclusion possible, but not as compelling. A stronger argument is that bookmaker betting limits may bind expert bettors from effectively exploiting this opportunity. These limits are put in place so that one bet cannot break the bookmaker. However, as the internet has substantially grown the size of the betting pool, bettors are able to have accounts with multiple betting sites all offering the same point spreads, implying that it is unlikely this constraint is binding. Additionally, why this constraint only affects this specific subset of games is unclear; that it would not have created market inefficiencies elsewhere is unexplained. Furthermore, in response to the four explanations listed above as well as all market inefficiency arguments, the existence of sports handicapping services provides an outlet for expert bettors to sell their information to novices. Quite popular services, they allow any bettor to follow the advice of betting industry experts. By selling expert information into the marketplace, adding to the wealth of knowledge the average bettor already has at his or her fingertips, it seems contradictory that a simple profitable betting rule could persist in the market. One legitimate problem with the above analysis is that it rests on the assumption that the bookmakers utilize the betting line to balance the amount of money bet on each side. This assumption seems sound as it is the stated goal of the company designing the betting lines and deviation from this strategy results in financial risk to the bookmaker. However in an experiment performed by Levitt (2004), he demonstrated that for the National Football League (NFL), bookmakers may be able to increase their profits by taking positions on the outcomes of football games. His data supported the theory that bookmakers are better able to predict both the

20 Jonathan Gibbs 20 May 11, 2007 outcomes of NFL games as well as bettors tendencies. This predictive superiority allows bookmakers to establish betting lines inducing uneven betting favoring the less likely outcome, increasing bookmaker profits despite the increased risk. However there are some issues with whether his conclusions hold when generalized away from his experiment. First, Levitt did not use market point spread and bettor s bets data. Instead, he used data taken from a contest offered through an online sports book where participants paid an upfront entry fee, picked the outcomes of five games each week, and were paid out at the end of the season corresponding to the participants rank in the number of games he or she picked correctly. Thus, bettors did not receive any direct payoff from correctly picking a specific game, only from cumulative number of wins. Also, as the payout structure of the contest was heavily skewed towards the top finishers, in order to maximize one s expected payout, there were incentives to pick contrary to one s competitors as opposed to picking the side more likely to win. Additionally, contestants could not express the relative intensity of preference over a game s outcome as all selections received equal weighting. Furthermore, the point spreads for each game were fixed at the announced point spread each Tuesday prior to the Sunday game. The set point spread feature prevented the market from adjusting the betting line throwing all of Levitt s results into question, particularly with reference to the paper by Gandar et al. Even if for arguments sake we believe that the assumption that the betting line is not designed to equalize the money bet on either side of the point spread for football, there are still further reasons why the assumption would hold for basketball. The major reason is that just as the two games are fundamentally different, the two betting markets are fundamentally different. Football games are concentrated predominantly on Sundays attracting many more bettors and many more betting dollars than basketball. The average NFL bettor tends to be less informed,

21 Jonathan Gibbs 21 May 11, 2007 betting instead to spice up a game or to bet on their home team, instead of systematically betting to make consistent profits. This difference is borne out by Paul and Weinbach s observation that the betting limit for an NBA game is approximately $2,200, less than half the $5,500 allowed per NFL game. This possibility is also consistent with Levitt s observation that if there is a large dollar amount bet outside his sample in the opposite direction of the bets within his sample, his sub sample conclusions would be incorrect. Since the betting line only demonstrates the median of the money bet and dollar volume on each side is the key amount, this issue could be problematic. Levitt also notes unlike NBA point spreads, NFL betting lines do not tend move, despite a week long trading period. One explanation for this difference is that the ease of scoring in football is greatly decreased as points are scored predominantly in 3s and 7s with final scores averaging in the 20s. In basketball with scores averaging in the high nineties and points scored in 1s, 2s, and 3s, it is much easier for a team to shift between covering the point spread and not covering the point spread. This ease discrepancy makes it more crucial for basketball point spreads to be accurate relative to football point spreads. Thus, despite possible evidence to the contrary for football, for basketball, it makes sense that the assumption of position neutrality on the part of the bookmaker would hold. Returning to the central question of how the NBA betting market can be dominated by informed bettors yet sustain simple profitable betting rules, this issue can be alternatively posited to ask the origin of the informed bettors superior abilities. Gandar et al. admit that it seems unlikely that bookmakers, armed with volumes of public data and with great monetary incentives to set accurate betting lines, incorrectly price point spreads with the frequency their analysis yielded. This contradiction signaled to Gandar et al. that informed traders either are better than the bookmakers at processing the public information into teams probabilities of covering or they

22 Jonathan Gibbs 22 May 11, 2007 possess superior private information. The first possibility does not seem credible as the bookmaker has at least as much if not many times more incentive to price point spreads accurately than does the bettor. Failure to price accurately and adjust betting lines exposes the bookmaker to financial risk and possibly financial ruin. However, the second option, that bettors possess superior private information over game outcomes bears strong resemblance to the illegal phenomenon of insider trading in financial markets. A heretofore unexplored aspect of the NBA betting market is whether players are using their unique superior knowledge over NBA teams and direct sphere of influence to affect game outcomes. To instead shift the attention to the National Collegiate Athletic Association s (NCAA) basketball betting market, the possibility that collegiate basketball players were impacting their play for direct monetary gain was studied in paper by Wolfers (2006). Prior to this paper, the possibility that point shaving was the cause of similar deviations from efficiency found in the college basketball point spread betting market 5, had never been raised. Wolfers similarly details the incentives for players to point shave, noting that the outcome which maximizes the joint surplus of the favored team s players and the gamblers is for the favored team to win but fail to cover the point spread. His analysis led to the startling conclusion that approximately 6% of games in which one team is favored by twelve points or more are influenced by point shaving. Given that games involving such strong favorites represent nearly 20% of the total population of games, his conclusions suggest about 1% of all college basketball games involve gambling related corruption. This paper analyzes the NBA betting market in a similar style to that done by Wolfers for the NCAA betting market. By attempting to demonstrate the existence of point shaving in NBA games, I hope to resolve conflict at the heart of this literature review; how informed bettors

23 Jonathan Gibbs 23 May 11, 2007 dominate the market, while published betting rules continue to garner profits. In addition, the data set used for this analysis includes information over in game scores. This data is new to the current literature and through its examination, this paper pushes further and studies the manner in which games conclude; specifically looking for evidence of decreased effort after the outcome is determined yet while there is still time left to affect the final margin. In this analysis I attempt to determine the incidence of point shaving in the NBA as well as unpack if and when it is happening. V. Data Set I: Data/Methodology In order to collect data and create a model to test for point shaving s possible incidence, I must first return to the point shaving decision theoretic framework presented in Section III. This framework detailed the underlying incentives to engage in point shaving as well as provided three critical aspects of its accomplishment. First, players choose to reduce their effort in order to not cover the betting line as this effort reduction is easily performed while the alternative of increasing one s effort may be exceedingly difficult or impossible. Second, favorites choose to shave points as the choice by an underdog to point shave implies a commitment to lose the game. Third, strong favorites can most easily shave points, as they are exposed to the least risk by maximizing the area where the favorite wins but fails to cover the point spread. Thus, if point shaving is occurring, it is in games where one team is favored by a large margin that we would expect to find distortionary effects. In trying to detect and measure these distortionary effects, I have created two data sets. The first data set, discussed in this section and the next, is used for general tests of the NBA betting market as well as an estimation of point shaving s prevalence. The second data set, discussed in Section VII and Section VIII closely examines the structure of the end of NBA games to see how outcomes relative to the point spread change in the final few

24 Jonathan Gibbs 24 May 11, 2007 minutes and whether the betting line has any explanatory power over the probability of the favored team covering. This first data set contains information over 15,859 NBA regular season games played between the and the season. For each game, the betting line and final score were recorded as well as unique game identifiers, such as the date and the teams that played. All information was taken from the available NBA final season logs freely provided at In their fiftieth year of publication, the Gold Sheet has the self proclaimed title of America s No. 1 Sports handicappers for over 4 decades, (The Gold Sheet 2007) their service being to provide statistics driven insight and edges into the sports betting world. The listed point spreads in the NBA final season logs are the consistent, agreed upon closing betting lines of the various online and casino sports books from which they draw their data. While postseason data was available in the logs, playoff games were excluded from the sample as being unnecessarily different in comparison to regular season games. In addition, there were 564 games for which there were no betting lines posted for reasons, such as uncertainty over a key player s participation. Games played at neutral sites were rare, and for simplicity were included as home and road games for the listed home and road teams. In full, this data set was assembled to test three aspects of the point shaving story. First, this data set contains more than twice the number of seasons and regular season games used in past literature studies. By examining this enlarged data set s basic characteristics, I wish to discover whether past conclusions regarding the betting market s efficiency still hold. Specifically, I use the log likelihood ratio test proposed by Even and Nobel (1992) and utilized by Paul and Weinbach (2005) to test for NBA gambling market efficiency and the profitability of individual betting rules. From the perspective of the underdog, the unrestricted log likelihood

25 Jonathan Gibbs 25 May 11, 2007 function takes the form L u, (3). The restricted form of the log likelihood function, L r, is derived from substituting (0.5) for (q hat ) in equation (3), as statistical market efficiency dictates that (q = 0.5). The formula to calculate the value of the likelihood ratio statistic for the efficient market null hypothesis of (q = 0.5) is found in (4). If a market is found to not be statistically efficient, the next step would be to test for the market s economic efficiency, or whether betting rules are profitable. To test this profitability null hypothesis, the eleven for ten trading rule dictates that (q = 0.524), as bettors must win more than 52.4% of their bets to offset the bookmakers take. This break even probability p is determined by: p + (1 p)*( 1.1) = 0 p = 1.1/2.1 p = The profitability test statistic value is calculated through the equation (5). Together, (4) and (5) test for statistical and economic NBA betting market efficiency while allowing for non normal residuals by not imposing equal mean or median restrictions on the forecast errors. L u = n[ln(q hat )] + (N n)[ln(1 q hat )] (3) 2(L u L r ) = 2{n[ln(q hat ) ln(0.5)] + (N n)[ln(1 q hat ) ln(0.5)]} (4) 2(L u L r ) = 2{n[ln(q hat ) ln(0.524)] + (N n)[ln(1 q hat ) ln(0.476)]} (5) N total number of observations n number of observations in which underdog covers q hat observed proportion of observations where the underdog covers The second use of this data set involves graphing kernel density functions in the mode of Figure 1, from Section III. Kernel density functions graph the probabilities of drawing specific values from a sample. The relevant statistic studied is the margin relative to the point spread, or the final margin normalized for the initial prediction. These graphs show the relative occurrence rates of betting line forecast errors. These errors are tested to see if they are symmetrically distributed about zero, indicating whether some point spread adjusted margins are more likely than others. By segmenting the data, these forecast errors are then examined to see if their